This paper describes an optimized training approach of radial basis function (RBF) classification by reducing a proposed classification-oriented error function. The training approach consists of two distinguished properties. First, radial basis functions, feature weights, and output weights can be updated iteratively; Second, it intrinsically distinguishes different learning contribution from training samples, which enables a large amount of learning from constructive samples, limited learning from outlier ones, and no learning at all from well trained ones.